CN111461159A - A Decoupled Representation Learning Algorithm Based on Similarity Constraints - Google Patents

Abstract

The invention discloses a decoupling representation learning algorithm based on similarity constraint, which aims to solve the problem of the existing InfoGAN in the unsupervised representation learning model. The specific steps of the present invention are as follows: step 1, prepare a data set; step 2, select the structure of the model: in order to generate a picture with better visual effect, the structure of WGAN using gradient penalty is adopted; step 3, impose similarity constraints on the model: in Based on the WGAN structure using gradient penalty, constraints are imposed on the similarity and factor of the generated images. The present invention proposes a similarity-constrained generative adversarial network (SCGAN). Compared with the InfoGAN model, the present invention has a simpler structure and only needs to add a similarity constraint on the basis of the original generative adversarial network. Meanwhile, SCGAN has a higher model Robustness, the performance is still stable in the processed dataset, and SCGAN is also an unsupervised representation learning model, so it can avoid expensive annotation work and has broad application prospects.

Description

A Decoupled Representation Learning Algorithm Based on Similarity Constraints

technical field

The invention relates to the field of picture representation, in particular to a decoupling representation learning algorithm based on similarity constraint.

Background technique

In probability and statistics theory, a generative model refers to a type of model that can estimate a probability distribution from training data and use this distribution to randomly generate new observation data. For an excellent machine learning algorithm to try to understand the inherent laws of data, it first needs to learn to create, that is to say, generating data is of great significance. Representation learning, as a hot field in generative learning, has received extensive attention from scholars. Efficient representations obtained through representation learning can assist many discriminative tasks in machine learning, such as classification, segmentation, detection, etc. Decoupled representation learning is a sub-branch of representation learning, which aims to learn factors that control the high-level semantic information of images. For supervised models, such as CGAN (Conditional Generative Adversarial Networks), the labels of the factors are provided explicitly, and the factors are learned to control the categories of items. For unsupervised models, such as InfoGAN (Information Maximization Generative Adversarial Network), mutual information is used to measure the relationship between factors and image representation, variational techniques are used to maximize the lower bound of mutual information, and factor learning is used to control the potential of the image. Representation, such as lighting, color, etc.

Conditional generative models need to provide labels for representation learning. In most cases, obtaining labels is expensive, and because conditional generative models provide labels for the representations that need to be learned, the representations captured by the model are limited, such as in Number type is captured on handwritten character set.

Among unsupervised representation learning models, InfoGAN is a classic. The idea of this model is to maximize the mutual information between the factor and the picture. The intuitive explanation is that since the factor can control a certain representation of the picture, the factor must have a close relationship with the picture, and mutual information can be used to measure this connection. However, the model of InfoGAN is more complicated. In order to maximize mutual information, InfoGAN uses variational technology and adds an additional neural network to maximize this lower bound. And the training of InfoGAN is not stable, and it is prone to collapse on some processed datasets (randomly translated handwritten character datasets), and people are also conducting related research.

SUMMARY OF THE INVENTION

The purpose of the embodiments of the present invention is to provide a decoupled representation learning algorithm based on similarity constraints, so as to solve the problems raised in the above background art.

To achieve the above purpose, the embodiments of the present invention provide the following technical solutions:

A decoupled representation learning algorithm based on similarity constraints, the specific steps are as follows:

Step 1, prepare the dataset;

Step 2: Choose the structure of the model: In order to generate better visual effects, the structure of WGAN (Wasserstein Distance-based Generative Adversarial Network) using gradient penalty is used. WGAN uses Wasserstein distance to replace the original GAN (Generation Adversarial Network). ), the Jensen-Shannon divergence in To allow the discriminator to evaluate the Wasserstein distance, it is necessary to constrain the discriminator under the 1-Lipschitzd (Lipschitz) limit, and the essence of the 1-Lipschitzd constraint is to require that the output of the discriminator change less than the input. The device approximates to the 1-Lipschitzd limit, such as weight clipping and gradient penalty. In practice, the effect of gradient penalty is much better than that of weight clipping, so gradient penalty technology is preferred here;

Step 3, impose similarity constraints on the model: On the basis of the WGAN structure using gradient penalty, constraints are imposed on the similarity and factors of the generated images. The core idea is that when the factors are the same, the representation difference of the generated images should be as small as possible. On the other hand, when the factors are different, the difference in the representation of the generated images should be as large as possible. The similarity constraint is to make different factors repel and attract the same factors, so that different factors can control different representations. achieve the purpose of expressing learning.

As a further solution of the embodiment of the present invention: in step 1, the data set includes simple data sets and complex data sets, which may have abundant samples, and provide the most comprehensive samples for the model to the greatest extent.

As a further solution of the embodiment of the present invention: the simple data set includes MNIST, Fashion-MNIST and SVHN, and the target is centered; the data set SVHN contains colorful digital pictures, the background color is richer, and each picture can appear with multiple numbers. Happening.

As a further solution of the embodiment of the present invention: the complex data set includes CIFAR-10 and CelebA, the data set CIFAR-10 contains pictures of real scenes, a total of 10 types of objects, and even objects of the same type may have huge differences.

As a further solution of the embodiment of the present invention: since the representation difference of the generated picture is not easy to obtain, it is proposed to assume that the generated picture is composed of content and representation. When the control content is the same, the representation difference of the generated picture is equivalent to the similarity of the generated picture. degree differences, thereby transforming the constraints from differences in generating image representations to differences in generating image similarities.

As a further solution of the embodiment of the present invention. The two datasets, MNIST and Fashion-MNIST, contain grayscale images, black backgrounds, and objects, with objects centered.

As a further solution of the embodiment of the present invention: the dataset CelebA contains a large number of star faces and has rich face representations.

Compared with the prior art, the beneficial effects of the embodiments of the present invention are:

The present invention proposes a Similarity Constrained Generative Adversarial Network (SCGAN). Compared with the InfoGAN model, the present invention has a simpler structure and only needs to add a similarity constraint on the basis of the original Generative Adversarial Network. Generative Adversarial Networks) have higher model robustness and are still stable in processed datasets, and SCGAN is also an unsupervised representation learning model, so it can avoid expensive labeling work and has broad application prospects.

Description of drawings

FIG. 1 is a work flow chart of Embodiment 1 of the similarity constraint-based decoupling representation learning algorithm.

Detailed ways

The technical solution of the present patent will be described in further detail below in conjunction with specific embodiments.

Example 1

A decoupled representation learning algorithm based on similarity constraints, the specific steps are as follows:

Step 1, randomly sample the noise z from the Gaussian distribution, and sample the factor c from the multinomial distribution or uniform distribution. The factor c can be a discrete factor or a continuous factor, and input the generator to generate a picture.

Step 2, randomly sample a batch of real images x from the data set and a batch of generated images G(z,c) from the generator output, and input the discriminator to let it distinguish between true and false.

Step 3, impose similarity constraints on the factor c and the generated image G(z,c).

Step 4, using the gradient descent algorithm to update the parameters of the discriminator and generator respectively.

Step 5, repeat steps 1-4 until the generator can generate enough realistic pictures, and different factors c can control different representations of the pictures.

The working principle of the embodiment of the present invention is that the method is based on a generative adversarial network and similarity constraints, and can learn decoupling representations unsupervised. Since it is an unsupervised learning method, there is no need to label the images in the dataset in advance. The representation of a picture itself is a relatively vague concept, and it is often difficult for people to identify a clear division boundary, such as the division of the brightness of the picture. The unsupervised method will automatically analyze the potential characteristics of the data set and learn some obvious changes in the image representation.

The model structure of this method is very simple, and it only needs to add similarity constraints on the basis of the original generative adversarial network. Another unsupervised model for decoupled representation learning, InfoGAN, needs to introduce an additional neural network to optimize the lower bound of mutual information, which increases the complexity of the model. And in terms of the difficulty of training, the model of this method is easier to converge, especially on some processed datasets (randomly translated handwritten character datasets), InfoGAN crashes during the training process, and this The model of the method is very robust.

The model of this method has better scalability. Since the similarity constraint only acts between the generated image and the factor, any novel GAN model can integrate our similarity constraint. In this way, the method can utilize these novel models to generate higher quality images, while the similarity constraints can ensure that the factor captures the changes in the image representation.

The above descriptions are only preferred embodiments of the present invention, and are not intended to limit the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the scope of the present invention. within the scope of protection. Any reference signs in the claims shall not be construed as limiting the involved claim.

In addition, it should be understood that although this specification is described in terms of embodiments, not each embodiment only includes an independent technical solution, and this description in the specification is only for the sake of clarity, and those skilled in the art should take the specification as a whole , the technical solutions in each embodiment can also be appropriately combined to form other implementations that can be understood by those skilled in the art.

Claims (7)

1. a decoupling representation learning algorithm based on similarity constraint, is characterized in that, concrete steps are as follows: Step 1, prepare the dataset; Step 2, select the structure of the model: adopt the structure of WGAN using gradient penalty; Step 3, impose similarity constraints on the model: On the basis of the WGAN structure using gradient penalty, constraints are imposed on the similarity and factors of the generated images. 2 . The similarity constraint-based decoupling representation learning algorithm according to claim 1 , wherein the data set in the first step includes a simple data set and a complex data set. 3 . 3. The similarity constraint-based decoupling representation learning algorithm according to claim 2, wherein the simple data set includes MNIST, Fashion-MNIST and SVHN. The decoupled representation learning algorithm based on similarity constraint according to claim 2 or 3, wherein the complex data set includes CIFAR-10 and CelebA. 5 . The decoupling representation learning algorithm based on similarity constraint according to claim 1 , wherein the generated picture is composed of two parts: content and representation. 6 . 6 . The decoupling representation learning algorithm based on similarity constraint according to claim 3 , wherein the two data sets of MNIST and Fashion-MNIST include grayscale images, black backgrounds and targets. 7 . 7 . The similarity constraint-based decoupling representation learning algorithm according to claim 4 , wherein the dataset CelebA contains a large number of star faces. 8 .

Images